Desired retention in FSRS, explained

If you read anything about FSRS, you will meet a number called desired retention, often 90 percent. It sounds technical, but it is really one simple trade-off: how much of what you review do you want to still remember when it comes back, and how many reviews are you willing to do to get there?

In short

• Desired retention is the probability of recalling a card at the moment you review it, as the scheduler aims for it.
• A higher target means shorter gaps and more reviews. A lower target means longer gaps, fewer reviews and more forgetting.
• At a target of 90 percent, the gap between reviews equals the card's stability, by the definition of stability.
• The cost of retention is not linear: each extra point near the top costs more reviews than the last.
• Retention is a workload decision, not a quality score. A lower number is not a worse learner.

What the number actually means

FSRS models each card with a memory state. One part of that state is stability, the number of days it takes for the chance of recall to fall from 100 percent to 90 percent. Another is retrievability, the probability that you can recall the card at this moment (the ABC of FSRS). If you have not read it, what is FSRS explains the model in more detail.

Right after a review, retrievability is high. As days pass, it falls. The scheduler has to decide how far to let it fall before it asks you again. Desired retention is that threshold.

Suppose a card has a stability of 20 days. By the definition of stability, retrievability falls to 90 percent after 20 days. So if your desired retention is 90 percent, the next review is scheduled about 20 days out. If you ask for 95 percent, the scheduler will not wait that long: it picks an earlier date, when the predicted chance of recall is still higher. If you ask for 80 percent, it waits longer.

Notice what this does and does not mean. A target of 90 percent does not mean you will remember exactly 90 percent of your library. It means that, on the day each card comes up, the model predicts a 90 percent chance you can recall it. Across many cards, the share you actually answer correctly should be near that. How near depends on how well the model fits you, which is why parameters fit to your own history matter.

The trade-off in plain terms

Think of retention as a dial between two kinds of cost.

• Turn it up and you forget less. You also review more often, because every card is brought back sooner. The daily queue grows.
• Turn it down and the queue shrinks. More cards will have faded a little by the time they appear, and more of your answers will be misses.

The FSRS documentation describes values between 70 percent and 97 percent as reasonable, and notes that higher retention means more daily reviews. Where the best value lies depends on what you are learning and how much time you have.

Why the extremes get expensive

Forgetting happens fast right after a review and slows afterward. That has an awkward consequence at the top of the range. To keep the chance of recall very high, the scheduler must review while the curve is still steep, which means short gaps. Moving from 85 to 90 percent costs some extra reviews; moving from 95 to 98 typically costs far more, for a gain you are unlikely to feel. At the other end, pushing the target far down saves reviews but makes recall unreliable, and relearning forgotten cards can cost more than reviewing them on time.

For that reason most people stay in the middle of the range.

How to think about choosing a value

There is no single correct number, but a few questions help.

1. What happens if you forget this? For a vocabulary list you will meet again in conversation, a miss is cheap, because it will come up again. For a drug interaction you must never get wrong, you may want a higher target for that material.
2. How much time do you really have? If you cannot sustain the daily queue a high target creates, the effective retention will be worse than a lower target you actually keep up with.
3. Is there a fixed deadline? Before an exam, higher retention for a short, defined period can be reasonable, with the plan to relax afterward.
4. How large is your deck? A small deck can afford a high target; a very large one multiplies the cost of every extra point.

A quiet trap: raising the target because misses feel bad. A miss is not a failure of the system. It is the point where the card gets more stable. Cards you never miss are being reviewed more than they need to be.

Retention and workload together

Workload in spaced repetition is roughly the number of cards in rotation divided by the typical gap between their reviews. Desired retention changes the typical gap. A higher target shrinks the gap for every card, so your queue per day grows in proportion across the whole library. That is why a small change can be felt on a large deck. It is also why adding new cards each day matters so much; how many flashcards per day walks through it.

Because the effect is cumulative, changes to the target are best made gradually and watched for a couple of weeks, not toggled day to day.

What about Cirello?

Cirello schedules with FSRS-6, and its parameters are retrained nightly on your review history. You never have to tune weights, and the default target of 90% works without touching anything. If you do want to change it, desired retention is a slider under Account - Study & scheduling: raise it for stronger recall with more reviews, lower it for fewer reviews with more forgetting. For how Cirello's nightly fitting works, see inside the FSRS-6 pipeline, and for the app-level picture, the FSRS app page.

A short checklist

• Retention is a threshold the scheduler aims for, not a grade on you.
• Higher costs more reviews, with sharply diminishing returns near the top.
• Pick a level you can sustain for months, not one that feels virtuous for a week.
• Fix the cause of misses with better cards before raising the target (how to write good flashcards).
• Change slowly, and judge the result over weeks.

The single most useful idea here is that remembering and reviewing trade against each other. Once you see the dial, a daily queue stops feeling like a verdict and starts feeling like a choice.